Masked Summarization to Generate Factually Inconsistent Summaries for Improved Factual Consistency Checking

TL;DR

This paper introduces a masked summarization technique to generate factually inconsistent summaries, enhancing the training of factual consistency classifiers and improving their alignment with human judgments.

Contribution

It proposes a novel method to generate factually inconsistent summaries by masking key information, aiding in better factual consistency checking.

Findings

Classifiers trained on generated summaries outperform existing models.

The method shows a strong correlation with human judgments.

Experiments on seven benchmark datasets validate effectiveness.

Abstract

Despite the recent advances in abstractive summarization systems, it is still difficult to determine whether a generated summary is factual consistent with the source text. To this end, the latest approach is to train a factual consistency classifier on factually consistent and inconsistent summaries. Luckily, the former is readily available as reference summaries in existing summarization datasets. However, generating the latter remains a challenge, as they need to be factually inconsistent, yet closely relevant to the source text to be effective. In this paper, we propose to generate factually inconsistent summaries using source texts and reference summaries with key information masked. Experiments on seven benchmark datasets demonstrate that factual consistency classifiers trained on summaries generated using our method generally outperform existing models and show a competitive correlation with human judgments. We also analyze the characteristics of the summaries generated using our method. We will release the pre-trained model and the code at https://github.com/hwanheelee1993/MFMA.